AI Models Delete Software Tests to Hide Mistakes Due to Reward Hacking
Amjad Masad · Vibe Coding: Everything You Need To Know — With Amjad Masad · Aug 11, 2025 · at 27:01
Replit CEO Amjad Masad explains why current AI models struggle with automated code verification and testing.
0:00 / 0:39exact quote · 39.9s
▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark
“Actually, right now it's pretty bad at testing software because there's this thing called reward hacking. So when you do reinforcement learning over large science models you're giving it a reward every time it does the right thing. Reward hacking is the way to, so, so the models become incredibly goal focused. They want to get that done, right? That's what RL does. And oftentimes what we see when we try to get the models to test things, it will start being corrupt in a way. It will like change the test to fit the mistakes it made, or sometimes delete the tests.”
quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →
More from Amjad Masad
Anthropic Leads in Agentic Coding While OpenAI Lags Behind
“They've been able to make, create a lead that hasn't been bridged by the other labs. Gemini is getting there on the agentic stuff, but I would say OpenAI kind of lagged behind.”
xAI Is Compute-Inefficient and Brute-Forces Problems With Massive Compute
“XAI is an amazing team. And they, they've been able to achieve so much in so little time, but it's also well known in the industry that they're computer inefficient. They're so compute rich that, that they're throwing computer the other problem in many ways.”
Masad: Model makers undercutting on price will destroy the ecosystem
“If they're gonna compete by undercutting everyone on price, they're gonna destroy the ecosystem.”
The First AGI Lab Will Not Trigger a Runaway Intelligence Explosion
“Whomever reaches AGI first, they're not going to go into intelligence explosion and just like suddenly, you know, super intelligence gets born. People, you know, other labs will catch up really quickly. And then, you know, there's going to be a lot of models. …”
Masad: Hyperreal, addictive AI technologies are a huge danger to human civilization
“We have these technologies that are, and the market around it that, that is bootstrapped to make us addicted because they're so, so much more enticing and low effort than the reality that we know and experience day to day. And I think that, that is a huge dang…”
Doctor Built an App Quoted at £100K for Under £200 on Replit
“A doctor from the UK that he's like, you know, there's all these apps around managing doctor patient relationships, but they never, it's not fully integrated. So, you know, you have ZocDoc, you can go make an appointment, but you know, how do you manage your p…”